QuantWM: training-free 2-bit KV cache quantization cuts video world model memory 6x without flicker
jiqizhixin · x · 2026-10-13
HIT Shenzhen's iLearn-Lab and NUS's LV-Lab present QuantWM, a training-free 2-bit KV cache quantization framework built specifically for autoregressive video world models.
The problem
- KV caches in video world models store every past frame for scene retrieval: generating just 5 seconds (93 frames) with LingBot-World-v2 needs over 21 GiB of KV-cache memory.
- Existing 2-bit quantization scores near BF16 on VBench for autoregressive video generation, but on world models the benchmark score barely moves while output flickers, blurs, and loses detail. Temporal stability, not average score, is the real failure mode.
Approach and results
- QuantWM analyzes why existing methods break temporal consistency — the cached historical frames are what the world model relies on for scene memory, and naive quantization damages that mechanism.
- Per the released figures, on HY-World-1.5 KV cache memory drops from 12.09 GiB (BF16) to 1.95 GiB (6x reduction) while VBench temporal-flickering score stays at 95.84% vs BF16's 95.62%, with visually consistent output.
More from Infra
- IBM moves Spyre accelerator to PyTorch Inductor, cutting compiler code 7x — PyTorch · 2026-10-13
- Your servers can report high availability while clients can't reach them — Franc0Fernand0 · 2026-10-13
- Model weights are only 1 of 4 VRAM allocations: why loaded LLMs still OOM — blaizedsouza · 2026-10-13
- US electricity chart: 4,079 TWh in 2025, data-center servers still a small slice of the pie — skorusARK · 2026-10-13
- While the US fights over data centers, China is building them fast — GeneReddit123 · 2026-10-13
- Scaledown's specialized models claim 200x faster, 467x cheaper inference with 1M-token context — cneuralnetwork · 2026-10-13